Using Predictor Variables


An adjustment predictor is a variable in the input data set that is used to adjust the forecast values produced by the forecasting model. Unlike a regressor, an adjustment variable does not have a regression coefficient. No model fitting is performed for adjustments. Nonmissing values of the adjustment series are simply added to the model prediction for the corresponding period. Missing adjustment values are ignored. If you supply adjustment values for observations within the period of fit, the adjustment values are subtracted from the actual values, and the model is fit to these adjusted values.

To add adjustments, select Add and then select Adjustments from the pop-up menu (shown in Figure 50.1). This displays the Adjustments Selection window. The Adjustments Selection window functions the same as the Regressor Selection window (which is shown in Figure 50.8). You can select any number of adjustment variables as predictors.

Unlike regressors, adjustments do not require forecasting models for the adjustment variables. If a variable that is used as an adjustment does have a forecasting model fit to it, the adjustment variable’s forecasting model is ignored when the variable is used as an adjustment.

You can use forecast adjustments to account for expected future events that have no precedent in the past and so cannot be modeled by regression. For example, suppose you are trying to forecast the sales of a product, and you know that a special promotional campaign for the product is planned during part of the period you want to forecast. If such sales promotion programs have been frequent in the past, then you can record the past and expected future level of promotional efforts in a variable in the data set and use that variable as a regressor in the forecasting model.

However, if this is the first sales promotion of its kind for this product, you have no way to estimate the effect of the promotion from past data. In this case, the best you can do is to make an educated guess at the effect the promotion will have and add that guess to what your forecasting model would predict in the absence of the special sales campaign.

Adjustments are also useful for making judgmental alterations to forecasts. For example, suppose you have produced forecast sales data for the next 12 months. Your supervisor believes that the forecasts are too optimistic near the end and asks you to prepare a forecast graph in which the numbers that you have forecast are reduced by 1000 in the last three months. You can accomplish this task by editing the input data set so that it contains observations for the actual data range of sales plus 12 additional observations for the forecast period, and a new variable called, for example, ADJUSTMENT. The variable ADJUSTMENT contains the value 1000 for the last three observations and is missing for all other observations. You fit the same model previously selected for forecasting by using the ARIMA Model Specification or Custom Model Specification window, but with an adjustment added that uses the variable ADJUSTMENT. Now when you graph the forecasts by using the Model Viewer, the last three periods of the forecast are reduced by 1000. The confidence limits are unchanged, which helps draw attention to the fact that the adjustments to the forecast deviate from what would be expected statistically.